5 min read
Always-on AI agents are coming to small business software, and the right number of jobs to give one this year is exactly one.
That is not a knock on the technology. It is a judgment about where the risk sits for a business with no IT department, no second pair of eyes and no one to call at 11 PM when something goes sideways. The newest wave of agents does not wait for you to type. It keeps working between conversations. That changes the math on mistakes, and solo owners should price that in before they get excited.
What “always-on” actually means now
According to launch coverage summarized by AI Agent Store’s weekly agent news roundup, OpenAI’s new Dots are agents that pursue goals you set even when you are not chatting with them. Each one reportedly runs on its own cloud computer, keeps a persistent memory of its objectives, and can research, create and edit files, and work with apps you connect. The initial rollout is to higher-tier paid plans rather than everyone.
OpenAI is not alone. Microsoft, Google and a growing list of startups are all pushing toward agents that run in the background. We have already looked at Copilot Autopilot and Amazon’s always-on seller assistant. The direction is clear: AI that works while you sleep.
The question is not whether this is impressive. It is. The question is what happens when it is wrong and you are not watching.
The case for restraint
Mistakes compound when nobody is looking
A chatbot that gives you a bad answer costs you one bad answer. You read it, frown, and move on. An agent that works for eight hours unsupervised can repeat the same misunderstanding a hundred times: a wrong price in every draft, a client misfiled in every note, a research thread chasing the wrong competitor all night. Autonomy multiplies whatever the agent believes, including its errors.
The security community is already worried
The OWASP Top 10 for large language model applications, a widely used security checklist for AI software, lists “excessive agency” as a core risk: giving an AI system more permissions, more tools or more independence than its task needs. That is a precise description of handing a background agent access to your inbox, your files and your payment tools all at once.
Large companies answer that risk with security teams, audit logs and approval workflows. A solo owner usually has a password manager and good intentions.
Your time savings are not guaranteed
Our look at the research on why a quarter of small businesses see no return from AI found the difference was rarely the tool. It was whether the owner defined a clear, narrow job. Always-on agents raise the stakes on that lesson. A vague goal given to a tireless worker produces a lot of vague output.
The strongest argument against me (and why it does not change the verdict)
The best counterargument is simple: solo owners are the people who most need help that does not require their attention. If you have to supervise an agent, the skeptic says, what is the point?
Fair. But the point of starting with one job is not to supervise forever. It is to learn how this particular agent fails on something low stakes, so you can decide what to trust it with next. Every serious adopter of new technology runs a pilot. A one-person business should run a smaller one, not skip it.
What a good “one job” looks like
Pick a task that scores well on all four of these:
- Read-heavy, write-light. Research, monitoring and summarizing beat sending, buying and deleting.
- Easy to check. You can verify the output in five minutes over coffee.
- Cheap when wrong. A mistake costs you a re-do, not a client or a refund.
- Repeating. It happens daily or weekly, so you get enough runs to judge reliability.
Good first jobs: a weekly scan of competitors’ public pricing pages, a morning summary of industry news in your niche, or a running list of local events you might sponsor. Poor first jobs: anything that emails clients directly, touches your bank or payment processor, or changes your website.
Guardrails worth setting on day one
- Least privilege. Connect only the apps the one job needs, and prefer read-only or draft-only permissions.
- A written brief. Two or three sentences describing the goal, what “done” looks like, and what the agent must never do.
- A daily check. Five minutes reviewing what it did yesterday. Keep a simple note of errors.
- A kill switch you know how to use. Before you start, find where to pause or delete the agent and disconnect its apps.
- A 30-day review. At the end of a month, decide based on your error notes, not on how impressive it felt.
If you want a more formal way to think about this, the US government’s NIST AI Risk Management Framework is free and boils down to a similar loop: know what the system does, measure how it performs, and manage the risks you find. You do not need to read all of it. The habit is what matters.
Where this leaves a solo owner in late 2026
Always-on agents are a real shift, and a year from now some of them will be dependable enough to run several parts of a small business. We are not there yet for most owners, and the vendors’ own staged rollouts suggest they know it. The same caution we argued for with AI video agents that can pass as human applies here: capability is arriving faster than the safety habits around it.
So give the agent one job. Watch it for a month. Then give it a second. That pace will feel slow next to the launch videos. It is also how you end up with a helper you trust instead of a cleanup project.
Which single, low-risk job would you hand an always-on agent first? If you cannot name one that passes all four tests above, that is your answer for now.
Related reading
- AI Agents Can Now Pay at Checkout. Small Sellers Should Fix the Basics First
- Microsoft Copilot Autopilot and Amazon’s Always-On Seller Assistant
- AI Saves Half of Small Businesses 4 Hours a Week. A Quarter See No Return
- AI Video Agents Can Now Pass as Human. That Is Exactly Why a Solo Business Should Wait
